Lots of scientific research is done on single sample, most of the work in CIA, which is my field, is done on single bottles of gas, no one I know of actually buys several and repeats the work. Single shot samples only mean you can't estimate the population variance independently for that attribute which is single sample, this isn't required for scientific work unless your goal is of course to obtain the population variance for that attribute.
In general you would bound the results by the QC of the manufacturer/maker, and thus unless they were horrible with huge defect rates, the result for one shot sampling would be expected to pass the standard significance values. To clearify, how many defects do you think a custom maker sends out of his shop? Lets assume he is fairly sloppy and messes up on heat treatment or grinds so badly that one out of 20 knives is significantly difference from standard.
Even this very high defect rate is not high enough to surpass the standard significance rate for confidence which is 5%, thus if all makers had similar you would expect 1/20 of the reviews to give skewed data, in general this holds for published work in any journal. It is also unlikely that this is actually close to the defect rate.
No work is 100% in that you are perfectly confident that what you see is actual relation, even with multiple samples. All it does is give you more confidence, it isn't a binary equation where one is right and the other wrong. At most you can say is something like there is a 3% chance this relationship is just random, which means 97% of the time what you propose is real, however it also means 3% of the time it isn't. No matter how many samples you do this can never be made 0/100%.
There are also lots of reviews with multiple trials, I have worked with many blades of the same style, and then you can cross correlated to other blades of the same geometry or steel to expand the sample and confirm results and then check with other reviewers, just like you would working in a lab. You just don't consider what you do in one piece of work but look for ways to reference everything else you have done and what other people have done, it doesn't have to be idential for this to be useful.
You also of course contact the maker/manufacturer and discuss expected/warrentied performance with them, and if you want most will provide multiple samples, but since they are hand picked you have to assume a bias. If you really want to get a solid performance of population variance you would get a decent sample size from multiple shipment dates from various shifts and so on.
But as noted, it isn't necessary to have a population variance for the information to be meaningful, it would of course make it more meaningful, and the more you had the more meaningful it would be.
-Cliff
